feat(data): add CutPaste-based synthetic anomaly generator with scar and union variants by geeky33 · Pull Request #3463 · open-edge-platform/anomalib

geeky33 · 2026-03-24T15:09:54Z

Description

Adds a CutPaste-based synthetic anomaly generator to the anomalib pipeline, extending the existing Perlin-based approach with feature-based anomaly synthesis.

Supports multiple variants (normal, scar, union) for improved diversity and realism of synthetic defects. The implementation is lightweight, CPU-friendly, and fully integrated into the existing dataset/datamodule pipeline with backward compatibility preserved.

🛠️ Fixes #3462

screenshots to support the PR:

✨ Changes

✅ Checklist

📚 I have made the necessary updates to the documentation (if applicable).
🧪 I have written tests that support my changes and prove that my fix is effective or my feature works (if applicable).
🏷️ My PR title follows conventional commit format.

…ehavior Signed-off-by: geeky33 <aaryap1204@gmail.com>

…and union variants Signed-off-by: geeky33 <aaryap1204@gmail.com> Made-with: Cursor

Copilot

Pull request overview

This PR adds a CutPaste-based synthetic anomaly generator to anomalib’s synthetic data pipeline (alongside the existing Perlin-based approach), and wires the new synthetic configuration options through datamodules, configs, and tests.

Changes:

Introduces a new CutPasteGenerator (normal/scar/union modes) and a visualization utility script.
Extends SyntheticAnomalyDataset/make_synthetic_dataset to support multiple generator backends via generator_type, plus configurable blend_factor, probability, and mask_threshold.
Refactors Perlin threshold-rescale logic into a shared helper and adds/updates unit tests and documentation/config snippets.

Reviewed changes

Copilot reviewed 30 out of 34 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tools/cutpaste_visualization.py	Adds a CLI utility to visualize CutPaste outputs and masks.
tests/unit/data/utils/test_synthetic.py	Adds tests for new synthetic dataset knobs and mask threshold behavior.
tests/unit/data/utils/test_perlin.py	Adds tests for shared Perlin utilities and GLASS compatibility helpers.
tests/unit/data/utils/test_cutpaste.py	Adds tests for the new CutPaste generator behaviors and modes.
tests/unit/data/datamodule/image/test_folder.py	Verifies datamodule forwards synthetic config into synthetic split creation.
src/anomalib/models/image/supersimplenet/anomaly_generator.py	Switches to shared Perlin threshold-rescale helper.
src/anomalib/models/image/dsr/anomaly_generator.py	Switches to shared Perlin threshold-rescale helper.
src/anomalib/data/utils/synthetic.py	Adds generator backend selection + config validation and change-mask computation.
src/anomalib/data/utils/generators/perlin.py	Adds shared Perlin rescale helper, GLASS defaults, and new configuration knobs.
src/anomalib/data/utils/generators/cutpaste.py	Implements the new CutPaste synthetic anomaly generator.
src/anomalib/data/utils/generators/init.py	Exposes CutPaste and new Perlin helper/constants via package exports.
src/anomalib/data/utils/init.py	Re-exports new generator utilities at `anomalib.data.utils` level.
src/anomalib/data/datamodules/image/visa.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/vad.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/tabular.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/realiad.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/mvtecad2.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/mvtecad.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/mvtec_loco.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/mpdd.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/kolektor.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/kaputt.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/folder.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/datumaro.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/btech.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/image/bmad.py	Forwards synthetic configuration args to the base datamodule.
src/anomalib/data/datamodules/base/image.py	Plumbs synthetic config into synthetic val/test split creation.
examples/configs/data/folder.yaml	Documents new synthetic configuration keys for the Folder datamodule example.
docs/source/snippets/config/data/image/folder/segmentation/cli/normal_and_synthetic.yaml	Updates CLI snippet to include new synthetic keys.
docs/source/snippets/config/data/image/folder/classification/cli/normal_and_synthetic.yaml	Updates CLI snippet to include new synthetic keys.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-03-24T15:14:08Z

+        return perlin_noise
+
+    denominator = perlin_noise.max() - perlin_noise.min()
+    if denominator == 0:


apply_perlin_threshold_rescale uses if denominator == 0: where denominator is a scalar tensor (perlin_noise.max() - perlin_noise.min()). This will raise RuntimeError: Boolean value of Tensor with more than one value is ambiguous (and will break the new unit test that passes uniform noise). Convert to a Python scalar (e.g., denominator.item()) or use a tensor-safe check like torch.isclose(denominator, torch.tensor(0., device=...)) before branching.

Suggested change

if denominator == 0:

if denominator.item() == 0:

Copilot · 2026-03-24T15:14:09Z

+    generator = CutPasteGenerator(
+        mode="normal",
+        enable_hflip=False,
+        enable_vflip=False,
+        enable_color_jitter=False,
+        rotation_range=(0.0, 0.0),
+        brightness_shift_range=(1.2, 1.2),
+    )
+    transformed_patch, _ = generator._transform_patch(patch)
+    assert not torch.allclose(transformed_patch, patch)


CutPasteGenerator._transform_patch now requires a selected_mode argument, but this test calls it with only the patch (generator._transform_patch(patch)). This will fail with a TypeError. Pass an explicit mode (e.g., 'normal') or update the helper signature to provide a default.

Signed-off-by: geeky33 <aaryap1204@gmail.com>

ashwinvaidya17 · 2026-03-26T09:51:37Z

@geeky33 thanks for the efforts. However before integrating the changes directly maybe it is better to start with a design document. I am concerned about changes to so many files. Can we come up with a design that makes the lest amount of changes to the existing files? Additionally, a point of consideration is how would new changes be integrated when we introduce more anomaly generation methods. Let's start with a document detailing stubs, examples, APIs etc

Copilot

Pull request overview

Copilot reviewed 30 out of 34 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

src/anomalib/data/utils/synthetic.py:119

The make_synthetic_dataset docstring still states that masks are generated by applying Perlin-noise perturbations, but the function now supports multiple backends via generator_type (Perlin and CutPaste). This description is misleading; update the narrative paragraph to describe the generic synthetic augmentation flow (and mention Perlin/CutPaste as options).

    """Convert normal samples into a mixed set with synthetic anomalies.

    The function generates synthetic anomalous images and their corresponding
    masks by applying Perlin noise-based perturbations to normal images.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-03-26T12:37:28Z

    if scale is None:
-        min_scale, max_scale = 0, 6
+        min_scale, max_scale = scale_exponent_range
        scalex = 2 ** torch.randint(min_scale, max_scale, (1,), device=device).item()
        scaley = 2 ** torch.randint(min_scale, max_scale, (1,), device=device).item()


generate_perlin_noise now accepts scale_exponent_range, but there is no validation that max_scale > min_scale. If a user passes an invalid range (e.g., equal bounds), torch.randint(min_scale, max_scale, ...) will raise a low-level error. Add an explicit check with a clear ValueError, and consider clarifying in the docstring that the upper bound is exclusive (since torch.randint samples from [min, max)).

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Signed-off-by: ayraa.ai <141430616+geeky33@users.noreply.github.com>

Copilot

Pull request overview

Copilot reviewed 30 out of 34 changed files in this pull request and generated 2 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-03-26T12:59:48Z

+
+        device = image.device
+        _, height, width = image.shape
+        if torch.rand(1, device=device).item() > self.probability:


CutPasteGenerator.generate applies augmentation when rand <= probability by returning early only on rand > probability. This makes probability=0.0 very rarely still apply (when rand is exactly 0), which can make the probability=0 unit test flaky. Use an apply-condition like rand < probability (or early-return on rand >= probability) to make boundary cases deterministic for 0.0 and 1.0.

Suggested change

if torch.rand(1, device=device).item() > self.probability:

if torch.rand(1, device=device).item() >= self.probability:

Copilot · 2026-03-26T12:59:49Z

+        raise ValueError(msg)
+    if generator_type == "cutpaste" and not isinstance(blend_factor, float):
+        msg = "For generator_type='cutpaste', blend_factor must be a float."
+        raise ValueError(msg)


_validate_synthetic_config requires blend_factor to be an instance of float for generator_type='cutpaste', which rejects valid numeric values like 0/1 (ints) that CutPasteGenerator can handle. Consider accepting any real scalar (e.g., numbers.Real) and adjusting the error message to clarify that CutPaste expects a single scalar (not a range tuple).

geeky33 and others added 3 commits March 21, 2026 04:37

refactor(data): expose synthetic blend factor and centralize perlin b…

89ea805

…ehavior Signed-off-by: geeky33 <aaryap1204@gmail.com>

Merge branch 'main' into anomaly

7534b30

Merge branch 'open-edge-platform:main' into anomaly

ba590db

Copilot AI review requested due to automatic review settings March 24, 2026 15:09

geeky33 requested review from ashwinvaidya17, rajeshgangireddy and samet-akcay as code owners March 24, 2026 15:09

Copilot started reviewing on behalf of geeky33 March 24, 2026 15:10 View session

feat(data): add CutPaste-based synthetic anomaly generator with scar …

2755222

…and union variants Signed-off-by: geeky33 <aaryap1204@gmail.com> Made-with: Cursor

geeky33 force-pushed the cutpaste branch from 6782890 to 2755222 Compare March 24, 2026 15:11

Copilot AI reviewed Mar 24, 2026

View reviewed changes

fix(prek): address cutpaste/synthetic ruff+mypy issues

5408f3d

Signed-off-by: geeky33 <aaryap1204@gmail.com>

ashwinvaidya17 added the Feature label Mar 26, 2026

Merge branch 'main' into cutpaste

701ba0a

Copilot AI review requested due to automatic review settings March 26, 2026 12:30

Copilot started reviewing on behalf of geeky33 March 26, 2026 12:31 View session

Copilot AI reviewed Mar 26, 2026

View reviewed changes

Update src/anomalib/data/utils/synthetic.py

5838d1d

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Signed-off-by: ayraa.ai <141430616+geeky33@users.noreply.github.com>

Copilot AI review requested due to automatic review settings March 26, 2026 12:54

Copilot started reviewing on behalf of geeky33 March 26, 2026 12:54 View session

Copilot AI reviewed Mar 26, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(data): add CutPaste-based synthetic anomaly generator with scar and union variants#3463

feat(data): add CutPaste-based synthetic anomaly generator with scar and union variants#3463
geeky33 wants to merge 7 commits into
open-edge-platform:mainfrom
geeky33:cutpaste

geeky33 commented Mar 24, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 24, 2026

Uh oh!

Copilot AI Mar 24, 2026

Uh oh!

ashwinvaidya17 commented Mar 26, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Copilot AI Mar 26, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 26, 2026

Uh oh!

Copilot AI Mar 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	if torch.rand(1, device=device).item() > self.probability:
	if torch.rand(1, device=device).item() >= self.probability:

Conversation

geeky33 commented Mar 24, 2026

Description

✨ Changes

✅ Checklist

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 24, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 24, 2026

Choose a reason for hiding this comment

Uh oh!

ashwinvaidya17 commented Mar 26, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Copilot AI Mar 26, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Mar 26, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 26, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants